Skip to content
体验新版
项目
组织
正在加载...
登录
切换导航
打开侧边栏
PaddlePaddle
PaddleDetection
提交
5008020d
P
PaddleDetection
项目概览
PaddlePaddle
/
PaddleDetection
大约 1 年 前同步成功
通知
695
Star
11112
Fork
2696
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
184
列表
看板
标记
里程碑
合并请求
40
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
PaddleDetection
项目概览
项目概览
详情
发布
仓库
仓库
文件
提交
分支
标签
贡献者
分支图
比较
Issue
184
Issue
184
列表
看板
标记
里程碑
合并请求
40
合并请求
40
Pages
分析
分析
仓库分析
DevOps
Wiki
0
Wiki
成员
成员
收起侧边栏
关闭侧边栏
动态
分支图
创建新Issue
提交
Issue看板
未验证
提交
5008020d
编写于
3月 20, 2018
作者:
武
武毅
提交者:
GitHub
3月 20, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #9154 from typhoonzero/pserver_parallel
pserver runs in parallel
上级
604a473a
d806fd41
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
99 addition
and
16 deletion
+99
-16
paddle/fluid/operators/listen_and_serv_op.cc
paddle/fluid/operators/listen_and_serv_op.cc
+34
-5
python/paddle/fluid/distribute_transpiler.py
python/paddle/fluid/distribute_transpiler.py
+65
-11
未找到文件。
paddle/fluid/operators/listen_and_serv_op.cc
浏览文件 @
5008020d
...
...
@@ -24,6 +24,7 @@ limitations under the License. */
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/proto_desc.h"
#include "paddle/fluid/framework/threadpool.h"
#include "paddle/fluid/operators/detail/grpc_server.h"
#include "paddle/fluid/operators/detail/sendrecvop_utils.h"
#include "paddle/fluid/operators/detail/simple_block_queue.h"
...
...
@@ -89,6 +90,10 @@ class ListenAndServOp : public framework::OperatorBase {
auto
*
block
=
Attr
<
framework
::
BlockDesc
*>
(
kOptimizeBlock
);
auto
*
program
=
block
->
Program
();
int
num_blocks
=
program
->
Size
();
PADDLE_ENFORCE_GE
(
num_blocks
,
2
,
"server program should have at least 2 blocks"
);
framework
::
Executor
executor
(
dev_place
);
// TODO(typhoonzero): change this to a while_op for every cluster-batch.
...
...
@@ -132,12 +137,36 @@ class ListenAndServOp : public framework::OperatorBase {
rpc_service_
->
ShutDown
();
break
;
}
try
{
executor
.
Run
(
*
program
,
&
recv_scope
,
block
->
ID
(),
/*global_block*/
false
/*create_local_scope*/
,
false
/*create_vars*/
);
}
catch
(
std
::
exception
&
e
)
{
LOG
(
ERROR
)
<<
"run sub program error "
<<
e
.
what
();
// put optimize blocks in the thread pool to start run, the last block
// should be global ops.
// NOTE: if is_gpu_place, CUDA kernels are laugched by multiple threads
// and this will still work.
std
::
vector
<
std
::
future
<
void
>>
fs
;
// block0 contains only listen_and_serv op, start run from block1.
for
(
int
blkid
=
1
;
blkid
<
num_blocks
-
1
;
++
blkid
)
{
fs
.
push_back
(
framework
::
Async
([
&
executor
,
&
program
,
&
recv_scope
,
blkid
]()
{
int
run_block
=
blkid
;
// thread local
try
{
executor
.
Run
(
*
program
,
&
recv_scope
,
run_block
,
false
/*create_local_scope*/
,
false
/*create_vars*/
);
}
catch
(
std
::
exception
&
e
)
{
LOG
(
ERROR
)
<<
"run sub program error "
<<
e
.
what
();
}
}));
}
for
(
int
i
=
0
;
i
<
num_blocks
-
2
;
++
i
)
fs
[
i
].
wait
();
// Run global block at final step, or block1 if there are only 2 blocks
if
(
num_blocks
>=
2
)
{
try
{
executor
.
Run
(
*
program
,
&
recv_scope
,
num_blocks
-
1
,
false
/*create_local_scope*/
,
false
/*create_vars*/
);
}
catch
(
std
::
exception
&
e
)
{
LOG
(
ERROR
)
<<
"run sub program error "
<<
e
.
what
();
}
}
// Reset the received sparse variables, the sum operator would not
// sum the input sparse variables which rows is empty at the next
// mini-batch.
...
...
python/paddle/fluid/distribute_transpiler.py
浏览文件 @
5008020d
...
...
@@ -307,15 +307,57 @@ class DistributeTranspiler:
# Iterate through the ops, and if an op and the optimize ops
# which located on current pserver are in one set, then
# append it into the sub program.
for
_
,
op
in
enumerate
(
self
.
optimize_ops
):
for
_
,
opt_op
in
enumerate
(
opt_op_on_pserver
):
if
ufind
.
is_connected
(
op
,
opt_op
):
if
self
.
_is_opt_op
(
op
):
self
.
_append_pserver_ops
(
optimize_block
,
op
,
endpoint
,
default_main_program
())
else
:
self
.
_append_pserver_non_opt_ops
(
optimize_block
,
op
)
break
# We try to put optimization program run parallelly, assume
# optimization program always looks like:
#
# prevop -> prevop -> opt op -> following op -> following op; ->
# prevop -> prevop -> opt op -> following op -> following op; ->
# global op -> global op
#
# we put operators that can run parallelly to many program blocks.
# in above example, we seperate ops by the ";". Global ops must run
# after all the optimize ops finished.
global_ops
=
[]
# HACK: optimization global ops only used to scale beta1 and beta2
# replace it with dependency engine.
for
op
in
self
.
optimize_ops
:
if
op
.
type
==
"scale"
:
for
in_name
in
op
.
input_arg_names
:
if
in_name
.
startswith
(
"beta1_pow_acc"
)
or
\
in_name
.
startswith
(
"beta2_pow_acc"
):
global_ops
.
append
(
op
)
def
__append_optimize_op__
(
op
,
block
):
if
self
.
_is_opt_op
(
op
):
self
.
_append_pserver_ops
(
block
,
op
,
endpoint
,
default_main_program
())
else
:
self
.
_append_pserver_non_opt_ops
(
block
,
op
)
# append op to the current block
per_opt_block
=
optimize_block
for
_
,
opt_op
in
enumerate
(
opt_op_on_pserver
):
for
_
,
op
in
enumerate
(
self
.
optimize_ops
):
# optimizer is connected to itself
if
ufind
.
is_connected
(
op
,
opt_op
)
and
\
op
not
in
global_ops
:
__append_optimize_op__
(
op
,
per_opt_block
)
per_opt_block
=
pserver_program
.
create_block
(
0
)
# append global ops
for
glb_op
in
global_ops
:
__append_optimize_op__
(
glb_op
,
per_opt_block
)
# NOT USED: single block version:
#
# for _, op in enumerate(self.optimize_ops):
# for _, opt_op in enumerate(opt_op_on_pserver):
# if ufind.is_connected(op, opt_op):
# __append_optimize_op__(glb_op, optimize_block)
# break
# step5 append the listen_and_serv op
pserver_program
.
global_block
().
append_op
(
type
=
"listen_and_serv"
,
...
...
@@ -660,10 +702,22 @@ class DistributeTranspiler:
# If one op's input is another op's output or
# one op's output is another op's input, we say
# the two operator is connected.
op1_input_names
=
op1
.
desc
.
input_arg_names
()
def
_append_inname_remove_beta
(
varname_list
):
op_input_names
=
[]
for
in_name
in
varname_list
:
# HACK: remove beta1 and beta2 to avoid let all
# ops connected.
if
in_name
.
startswith
(
"beta2_pow_acc"
)
or
\
in_name
.
startswith
(
"beta1_pow_acc"
):
continue
else
:
op_input_names
.
append
(
in_name
)
return
op_input_names
op1_input_names
=
_append_inname_remove_beta
(
op1
.
desc
.
input_arg_names
())
op1_output_names
=
op1
.
desc
.
output_arg_names
()
op2_input_names
=
op2
.
desc
.
input_arg_names
(
)
op2_input_names
=
_append_inname_remove_beta
(
op2
.
desc
.
input_arg_names
()
)
op2_output_names
=
op2
.
desc
.
output_arg_names
()
if
set
(
op1_output_names
)
&
set
(
op2_input_names
)
or
\
...
...
编辑
预览
Markdown
is supported
0%
请重试
或
添加新附件
.
添加附件
取消
You are about to add
0
people
to the discussion. Proceed with caution.
先完成此消息的编辑!
取消
想要评论请
注册
或
登录